{"id":4667,"date":"2026-05-11T15:00:25","date_gmt":"2026-05-11T07:00:25","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/05\/11\/%e5%85%b7%e8%ba%ab%e5%a4%a7%e6%a8%a1%e5%9e%8br1%e6%97%b6%e5%88%bb%ef%bc%9alibero%e7%bb%88%e7%bb%93%e8%80%85%ef%bc%8c99-9%e8%83%8c%e5%90%8e%e7%9a%84%e7%89%a9%e7%90%86%e6%8e%a8%e7%90%86%e6%96%b0\/"},"modified":"2026-05-11T15:00:25","modified_gmt":"2026-05-11T07:00:25","slug":"%e5%85%b7%e8%ba%ab%e5%a4%a7%e6%a8%a1%e5%9e%8br1%e6%97%b6%e5%88%bb%ef%bc%9alibero%e7%bb%88%e7%bb%93%e8%80%85%ef%bc%8c99-9%e8%83%8c%e5%90%8e%e7%9a%84%e7%89%a9%e7%90%86%e6%8e%a8%e7%90%86%e6%96%b0","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/05\/11\/%e5%85%b7%e8%ba%ab%e5%a4%a7%e6%a8%a1%e5%9e%8br1%e6%97%b6%e5%88%bb%ef%bc%9alibero%e7%bb%88%e7%bb%93%e8%80%85%ef%bc%8c99-9%e8%83%8c%e5%90%8e%e7%9a%84%e7%89%a9%e7%90%86%e6%8e%a8%e7%90%86%e6%96%b0\/","title":{"rendered":"\u5177\u8eab\u5927\u6a21\u578bR1\u65f6\u523b\uff1aLIBERO\u7ec8\u7ed3\u8005\uff0c99.9%\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u65b0\u8303\u5f0f"},"content":{"rendered":"
\"\u5177\u8eab\u5927\u6a21\u578bR1\u65f6\u523b\uff1aLIBERO\u7ec8\u7ed3\u8005\uff0c99.9%\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u65b0\u8303\u5f0f\"\/<\/div><\/figure>\n
\"\u5177\u8eab\u5927\u6a21\u578bR1\u65f6\u523b\uff1aLIBERO\u7ec8\u7ed3\u8005\uff0c99.9%\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u65b0\u8303\u5f0f\"\/<\/div><\/figure>\n
\"\u5177\u8eab\u5927\u6a21\u578bR1\u65f6\u523b\uff1aLIBERO\u7ec8\u7ed3\u8005\uff0c99.9%\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u65b0\u8303\u5f0f\"\/<\/div><\/figure>\n
\"\u5177\u8eab\u5927\u6a21\u578bR1\u65f6\u523b\uff1aLIBERO\u7ec8\u7ed3\u8005\uff0c99.9%\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u65b0\u8303\u5f0f\"\/<\/div><\/figure>\n

\u5141\u4e2d \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u673a\u5668\u4eba\u62c9\u4e2a\u62c9\u94fe\uff0c\u5230\u5e95\u9700\u4e0d\u9700\u8981\u201c\u8111\u5b50\u201d\uff1f<\/p>\n

\u8fc7\u53bb\u51e0\u5e74\uff0c\u4eceOpenVLA\u5230\u03c00\u3001\u03c00.5\uff0c\u5177\u8eab\u5927\u6a21\u578b\u5df2\u7ecf\u80fd\u8ba9\u673a\u5668\u4eba\u628a\u6307\u4ee4\u548c\u52a8\u4f5c\u8fde\u5f97\u6709\u6a21\u6709\u6837\u3002<\/p>\n

\u4f46\u4e00\u65e6\u5305\u7684\u4f4d\u7f6e\u632a\u4e86\u51e0\u5398\u7c73\uff0c\u6216\u8005\u5149\u7167\u6697\u4e86\u4e00\u70b9\uff0c\u5b83\u4eec\u5f80\u5f80\u5c31\u4f1a\u201c\u5927\u8111\u5b95\u673a\u201d\u3002<\/p>\n

\u7a76\u5176\u539f\u56e0\uff0c\u662f\u56e0\u4e3a\u8fd9\u4e9b\u673a\u5668\u4eba\u5927\u591a\u5728\u73a9\u201c\u8fde\u8fde\u770b\u201d\uff1a\u770b\u5230\u89c2\u5bdf\u7ed3\u679c\uff0c\u76f4\u63a5\u8f93\u51fa\u52a8\u4f5c\u3002<\/p>\n

\u5b83\u4eec\u53ea\u662f\u8bb0\u4f4f\u4e86\u8f68\u8ff9\uff0c\u5374\u5e76\u4e0d\u7406\u89e3\u80cc\u540e\u7684\u7269\u7406\u903b\u8f91\u3002<\/p>\n

\u73b0\u5728\uff0c\u4e00\u79cd\u8ba9\u673a\u5668\u4eba\u201c\u5148\u60f3\u660e\u767d\uff0c\u518d\u7a33\u5b9a\u884c\u52a8\u201d\u7684\u65b0\u8303\u5f0f\u6765\u4e86\u3002<\/p>\n

\u7531\u81f3\u7b80\u52a8\u529b\u3001\u5317\u5927\u3001\u6e2f\u4e2d\u6587\u8054\u5408\u63d0\u51fa\u7684LaST-R1\uff0c\u9996\u6b21\u5c06\u9690\u7a7a\u95f4\u7269\u7406\u63a8\u7406\u585e\u8fdb\u4e86\u5f3a\u5316\u5b66\u4e60\u7684\u95ed\u73af\u3002<\/p>\n

\u540c\u65f6\uff0cLaST-R1\u4f5c\u4e3aLaST\u2080\u57fa\u5ea7\u6a21\u578b\u7684\u7269\u7406\u4e16\u754c\u540e\u8bad\u7ec3\u8303\u5f0f\uff0cLaST\u2080\u9996\u521b\u9762\u5411\u673a\u5668\u4eba\u7684\u9690\u7a7a\u95f4\u7269\u7406\u601d\u7ef4\u94fe\u63a8\u7406\uff0c\u5e76\u5df2\u4e2d\u7a3fICML2026 Spotlight\uff08top 2.2%\uff09\u3002<\/p>\n

\u5b83\u7684\u8868\u73b0\u6709\u591a\u5938\u5f20\uff1f<\/p>\n

\u8fd9\u4e2a\u8ba9\u673a\u5668\u4eba\u957f\u51fa\u201c\u7269\u7406\u8111\u201d\u7684LaST-R1\uff0c\u5230\u5e95\u662f\u600e\u4e48\u70bc\u6210\u7684\uff1f<\/p>\n

\u90a3\u4e2a\u8ba9\u73af\u5883\u53cd\u9988\u540c\u65f6\u4f18\u5316\u201c\u600e\u4e48\u60f3\u201d\u548c\u201c\u600e\u4e48\u52a8\u201d\u7684LAPO\u7b97\u6cd5\u53c8\u85cf\u7740\u4ec0\u4e48\u7384\u673a\uff1f<\/p>\n

\u6211\u4eec\u987a\u7740\u8fd9\u7bc7\u8bba\u6587\uff0c\u6df1\u6316\u4e86\u4e00\u4e0b\u8fd9\u5957\u80fd\u8ba9\u673a\u5668\u4eba\u201c\u6df1\u601d\u719f\u8651\u201d\u7684\u540e\u8bad\u7ec3\u9ed1\u79d1\u6280\u3002<\/p>\n

\u5c3d\u7ba1\u4eceOpenVLA\u5230\u03c00.5\uff0c\u5177\u8eab\u5927\u6a21\u578b\u5df2\u7ecf\u5b8c\u6210\u4e86\u56fe\u50cf\u3001\u8bed\u8a00\u4e0e\u52a8\u4f5c\u7684\u521d\u6b65\u5bf9\u9f50\u3002<\/p>\n

\u4f46\u5728\u5b9e\u9645\u843d\u5730\u4e2d\uff0c\u5de5\u4e1a\u754c\u53d1\u73b0\u4e86\u4e00\u4e2a\u81f4\u547d\u7684\u201c\u5e7b\u89c9\u201d\uff1a<\/p>\n

\u80fd\u6a21\u4eff\uff0c\u4e0d\u7b49\u4e8e\u80fd\u5728\u7269\u7406\u4e16\u754c\u6cdb\u5316\u3002<\/p>\n

\u8fd9\u5c31\u5bfc\u81f4\u4e86\u6781\u5dee\u7684\u6cdb\u5316\u6027\u3002<\/p>\n

\u6253\u4e2a\u6bd4\u65b9\uff0c\u673a\u5668\u4eba\u53ef\u80fd\u8bb0\u4f4f\u4e86100\u79cd\u62c9\u62c9\u94fe\u7684\u8f68\u8ff9\uff0c\u4f46\u53ea\u8981\u62c9\u94fe\u7684\u89d2\u5ea6\u504f\u8f6c15\u5ea6\uff0c\u6216\u8005\u5149\u7167\u53d1\u751f\u53d8\u5316\uff0c\u5355\u7eaf\u9760\u201c\u89c2\u5bdf\u2192\u52a8\u4f5c\u201d\u7684\u7aef\u5230\u7aef\u6620\u5c04\u5c31\u4f1a\u5931\u6548\u3002<\/p>\n

\u6838\u5fc3\u95ee\u9898\u5728\u4e8e\uff0c\u73b0\u6709\u7684VLA\u6a21\u578b\u7f3a\u5c11\u4e00\u4e2a\u201c\u601d\u8003\u201d\u7684\u4e2d\u95f4\u5c42\u2014\u2014\u5373\u8ba9\u673a\u5668\u4eba\u5728\u884c\u52a8\u4e4b\u524d\uff0c\u5bf9\u7269\u7406\u4e16\u754c\u8fdb\u884c\u63a8\u7406\u3002<\/p>\n

\u8fc7\u53bb\uff0c\u5b66\u672f\u754c\u4e5f\u66fe\u5c1d\u8bd5\u5f15\u5165\u601d\u7ef4\u94fe\uff08CoT\uff09\u6765\u89e3\u51b3\u63a8\u7406\u95ee\u9898\u3002<\/p>\n

\u4f46\u5bf9\u4e8e\u673a\u5668\u4eba\u64cd\u4f5c\u800c\u8a00\uff0c\u8bed\u8a00\u63a8\u7406\u5f80\u5f80\u592a\u6162\u4e14\u9897\u7c92\u5ea6\u592a\u7c97\uff0c\u4f60\u5f88\u96be\u7528\u6587\u5b57\u7cbe\u51c6\u63cf\u8ff0\u201c\u62c9\u94fe\u54ac\u5408\u65f6\u7684\u7ec6\u5fae\u963b\u529b\u53cd\u9988\u201d\u3002<\/p>\n

LaST-R1\u7684\u6838\u5fc3\u7a81\u7834\uff0c\u5c31\u662f\u653e\u5f03\u4e86\u4f4e\u6548\u7684\u8bed\u8a00CoT\uff0c\u8f6c\u800c\u5728\u9690\u7a7a\u95f4\uff08Latent Space\uff09\u4e2d\u6784\u5efa\u7269\u7406\u63a8\u7406\u94fe\u3002<\/p>\n

\u5b83\u4e0d\u518d\u8ba9\u673a\u5668\u4eba\u770b\u5230\u56fe\u50cf\u5c31\u201c\u95ed\u773c\u201d\u51fa\u52a8\u4f5c\uff0c\u800c\u662f\u5148\u5728\u9690\u6027\u7a7a\u95f4\u91cc\u5efa\u6a21\u573a\u666f\u7684\u7ed3\u6784\u3001\u7269\u4f53\u7684\u7269\u7406\u5173\u7cfb\u4ee5\u53ca\u672a\u6765\u7684\u52a8\u6001\u53d8\u5316\u3002<\/p>\n

\u7136\u800c\uff0c\u8981\u8ba9\u673a\u5668\u4eba\u5b66\u4f1a\u8fd9\u79cd\u201c\u601d\u8003\u201d\uff0c\u4ec5\u9760\u9759\u6001\u7684\u6a21\u4eff\u5b66\u4e60\uff08SFT\uff09\u662f\u4e0d\u591f\u7684\u3002<\/p>\n

\u76ee\u524d\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u65b9\u6cd5\u5927\u591a\u50cf\u662f\u4e00\u4e2a\u53ea\u770b\u7ed3\u679c\u7684\u4e25\u5389\u6559\u7ec3\uff1a\u5b83\u53ea\u544a\u8bc9\u673a\u5668\u4eba\u52a8\u4f5c\u6210\u6ca1\u6210\u529f\uff08\u4f18\u5316Action Space\uff09\uff0c\u5374\u65e0\u6cd5\u6307\u5bfc\u673a\u5668\u4eba\u201c\u521a\u624d\u90a3\u4e0b\u4f60\u662f\u600e\u4e48\u60f3\u7684\u201d\u3002<\/p>\n

\u9488\u5bf9\u8fd9\u4e00\u75db\u70b9\uff0c\u8be5\u56e2\u961f\u63d0\u51fa\u7684LAPO\uff08Latent-to-Action Policy Optimization\uff09\u7b97\u6cd5\uff0c\u6b63\u5f0f\u5c06\u201c\u601d\u8003\u8fc7\u7a0b\u201d\u62c9\u8fdb\u4e86\u5f3a\u5316\u5b66\u4e60\u7684\u4f18\u5316\u95ed\u73af\u3002<\/p>\n

\u5b83\u8ba9\u73af\u5883\u53cd\u9988\u4e0d\u4ec5\u4f18\u5316\u52a8\u4f5c\uff0c\u4e5f\u4f18\u5316\u673a\u5668\u4eba\u884c\u52a8\u524d\u7684\u201c\u7269\u7406\u601d\u8003\u201d\u3002<\/p>\n

\u8fd1\u65e5\uff0c\u81f3\u7b80\u52a8\u529b\u3001\u9999\u6e2f\u4e2d\u6587\u5927\u5b66\u3001\u5317\u4eac\u5927\u5b66\u8ba1\u7b97\u673a\u5b66\u9662\u591a\u5a92\u4f53\u4fe1\u606f\u5904\u7406\u56fd\u5bb6\u91cd\u70b9\u5b9e\u9a8c\u5ba4\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u9762\u5411\u673a\u5668\u4eba\u64cd\u4f5c\u7684\u81ea\u9002\u5e94\u7269\u7406\u9690\u7a7a\u95f4\u63a8\u7406\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u2014\u2014<\/p>\n

LaST-R1\uff08Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning\uff09\u3002<\/p>\n

\u5b83\u5e0c\u671b\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u540e\u8bad\u7ec3\uff0c\u8ba9\u5177\u8eab\u5927\u6a21\u578b\u4e0d\u4ec5\u5b66\u4f1a\u751f\u6210\u52a8\u4f5c\uff0c\u4e5f\u5b66\u4f1a\u5728\u884c\u52a8\u524d\uff0c\u8fdb\u884c\u9762\u5411\u7269\u7406\u4e16\u754c\u7684\u9690\u7a7a\u95f4\u63a8\u7406\u3002<\/p>\n

\u4e0e\u4ee5\u5f80\u4e3b\u8981\u4f18\u5316action space\uff08\u52a8\u4f5c\u7a7a\u95f4\uff0c\u5373\u673a\u5668\u4eba\u6240\u6709\u53ef\u6267\u884c\u6307\u4ee4\u7684\u96c6\u5408\uff09\u7684\u5177\u8eab\u5927\u6a21\u578bRL\u4e0d\u540c\uff0cLaST-R1\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a<\/p>\n

\u673a\u5668\u4eba\u4e0d\u5e94\u53ea\u4ece\u56fe\u50cf\u548c\u6307\u4ee4\u76f4\u63a5\u9884\u6d4b\u4e0b\u4e00\u6b65\u52a8\u4f5c\uff0c\u800c\u5e94\u5148\u5728latent space\uff08\u9690\u7a7a\u95f4\uff0c\u53ef\u4ee5\u7406\u89e3\u4e3a\u673a\u5668\u4eba\u5927\u8111\u91cc\u7684\u201c\u9690\u6027\u8ba4\u77e5\u5c42\u201d\uff09\u4e2d\u7406\u89e3\u573a\u666f\u7ed3\u6784\u3001\u7269\u4f53\u5173\u7cfb\u548c\u7269\u7406\u52a8\u6001\uff0c\u518d\u751f\u6210\u66f4\u7a33\u5b9a\u3001\u7cbe\u51c6\u7684\u52a8\u4f5c\u3002<\/p>\n

\u6362\u53e5\u8bdd\u8bf4\uff1a<\/p>\n

LaST-R1\u4e0d\u53ea\u4f18\u5316\u673a\u5668\u4eba\u7684\u201c\u624b\u201d\uff0c\u4e5f\u4f18\u5316\u5b83\u7684\u201c\u8111\u201d\u3002<\/p>\n

\u5177\u4f53\u6765\u770b\uff0cLaST-R1\u6784\u5efa\u4e86\u4e00\u4e2a\u9762\u5411latent reasoning-before-acting\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u540e\u8bad\u7ec3\u6846\u67b6\uff0c\u6838\u5fc3\u7531\u4e09\u6b65\u7ec4\u6210\uff1a<\/p>\n

1\u3001\u7269\u7406\u9690\u7a7a\u95f4\u63a8\u7406\u5efa\u6a21\uff08Physical Latent Reasoning\uff09<\/p>\n

2\u3001\u9690\u7a7a\u95f4\u63a8\u7406\u4e0e\u52a8\u4f5c\u751f\u6210\u7684\u8054\u5408\u5f3a\u5316\u4f18\u5316\uff08Latent-to-Action Policy Optimization\uff09<\/p>\n

3\u3001\u81ea\u9002\u5e94latent CoT\u63a8\u7406\u673a\u5236\uff08Adaptive Latent CoT\uff09<\/p>\n

LaST-R1\u6539\u53d8\u7684\u662f\u5177\u8eab\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u4f18\u5316\u5bf9\u8c61\uff1a\u4ece\u53ea\u4f18\u5316\u52a8\u4f5c\uff0c\u8f6c\u5411\u540c\u65f6\u4f18\u5316\u52a8\u4f5c\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u3002<\/p>\n

\u7814\u7a76\u56e2\u961f\u5728\u4eff\u771f\u548c\u771f\u673a\u73af\u5883\u4e2d\u90fd\u8fdb\u884c\u4e86\u7cfb\u7edf\u9a8c\u8bc1\u3002<\/p>\n

\u5728\u4eff\u771fLIBERO benchmark\u4e0a\uff0cLaST-R1\u4ec5\u4f9d\u8d561\u6761\u8f68\u8ff9\u5b8c\u6210warm-up\uff0c\u968f\u540e\u901a\u8fc7\u5728\u7ebfRL\u4f18\u5316\uff0c\u6700\u7ec8\u53d6\u5f9799.9%\u5e73\u5747\u6210\u529f\u7387\uff0c\u5e76\u76f8\u6bd4Action-Only+PPO\u5c55\u73b0\u51fa\u66f4\u5feb\u6536\u655b\u548c\u66f4\u9ad8\u6700\u7ec8\u6027\u80fd\u3002<\/p>\n

\u5728\u771f\u673a\u90e8\u7f72\u4e2d\uff0cLaST-R1\u4ec5\u4f7f\u752830\u6761\u8f68\u8ff9warm-up\uff0c\u518d\u901a\u8fc7RL\u540e\u8bad\u7ec3\u5c06\u5e73\u5747\u6210\u529f\u7387\u4ece52.5%\u63d0\u5347\u523093.75%\uff0c\u663e\u8457\u8d85\u8fc7\u4f7f\u7528100\u6761\u4e13\u5bb6\u8f68\u8ff9\u7684\u03c00.5\uff0871.25%\uff09\u3002<\/p>\n

\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5728\u771f\u5b9e\u6270\u52a8\u6761\u4ef6\u4e0b\uff0cLaST-R1\u4ecd\u4fdd\u6301\u8f83\u5c0f\u6027\u80fd\u4e0b\u964d\uff0c\u8bf4\u660e\u5176\u5b66\u4e60\u5230\u7684\u4e0d\u662f\u5355\u4e00\u573a\u666f\u4e2d\u7684\u52a8\u4f5c\u8f68\u8ff9\uff0c\u800c\u662f\u66f4\u53ef\u8fc1\u79fb\u7684\u7a7a\u95f4\u8bed\u4e49\u548c\u7269\u7406\u52a8\u6001\u7406\u89e3\u3002<\/p>\n

\u4e0a\u8ff0\u7ed3\u679c\u610f\u5473\u7740\uff0c\u5177\u8eab\u5927\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\u7684\u91cd\u70b9\u6b63\u5728\u53d1\u751f\u53d8\u5316\u2014\u2014<\/p>\n

\u673a\u5668\u4eba\u4e0d\u518d\u53ea\u662f\u901a\u8fc7RL\u5b66\u4f1a\u66f4\u719f\u7ec3\u5730\u6267\u884c\u52a8\u4f5c\uff0c\u800c\u662f\u5f00\u59cb\u901a\u8fc7RL\u5b66\u4f1a\u66f4\u5408\u7406\u5730\u8fdb\u884c\u7269\u7406\u63a8\u7406\u3002<\/p>\n

LaST-R1\u7684\u610f\u4e49\uff0c\u5728\u4e8e\u5b83\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u5177\u8eab\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u8303\u5f0f\uff0c\u80fd\u591f\u8ba9\u73af\u5883\u53cd\u9988\u540c\u65f6\u5851\u9020\u673a\u5668\u4eba\u7684\u201c\u601d\u8003\u65b9\u5f0f\u201d\u548c\u201c\u884c\u52a8\u65b9\u5f0f\u201d\u3002<\/p>\n

\u4e00\u65e6\u9690\u7a7a\u95f4\u63a8\u7406\u4ece\u6a21\u4eff\u5b66\u4e60\u7684\u201c\u9759\u6001\u811a\u672c\u201d\u8fdb\u5316\u4e3a\u5f3a\u5316\u5b66\u4e60\u7684\u201c\u6f14\u8fdb\u6838\u5fc3\u201d\uff0c\u673a\u5668\u4eba\u4fbf\u80fd\u6446\u8131\u5bf9\u6f14\u793a\u6570\u636e\u7684\u523b\u677f\u590d\u73b0\u3002<\/p>\n

\u5728\u4e0d\u65ad\u7684\u4ea4\u4e92\u8bd5\u9519\u4e2d\uff0c\u5b83\u4eec\u5f00\u59cb\u5f3a\u5316\u6a21\u578b\u7684\u7269\u7406\u63a8\u7406\u3002<\/p>\n

\u8fd9\u6216\u8bb8\u4e5f\u662f\u5177\u8eab\u5927\u6a21\u578b\u4ece\u201c\u4f1a\u6a21\u4eff\u201d\u8d70\u5411\u201c\u4f1a\u9002\u5e94\u201d\u7684\u5173\u952e\u4e00\u6b65\u3002<\/p>\n

\u6574\u4e2aLaST-R1\u6846\u67b6\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e09\u4e2a\u5173\u952e\u9636\u6bb5\uff1a\u5148\u63a8\u7406\u3001\u518d\u4f18\u5316\u3001\u52a8\u6001\u51b3\u5b9a\u60f3\u591a\u4e45\u3002<\/p>\n

\u7b2c\u4e00\u9636\u6bb5\uff1a\u884c\u52a8\u524d\u7684\u9690\u7a7a\u95f4\u63a8\u7406\uff08Latent Reasoning-before-Acting\uff09<\/p>\n

\u7ed9\u5b9a\u5f53\u524d\u89c6\u89c9\u89c2\u6d4b\u548c\u8bed\u8a00\u6307\u4ee4\uff0cLaST-R1\u4e0d\u4f1a\u76f4\u63a5\u751f\u6210\u52a8\u4f5c\uff0c\u800c\u662f\u5148\u751f\u6210\u4e00\u6bb5\u9690\u7a7a\u95f4\u63a8\u7406\u5d4c\u5165\uff08latent reasoning embeddings\uff09\uff0c\u4f5c\u4e3a\u884c\u52a8\u524d\u7684\u201c\u9690\u7a7a\u95f4\u7269\u7406\u601d\u8003\u201d\uff0c\u7528\u4e8e\u5efa\u6a21\u7269\u4f53\u5173\u7cfb\u3001\u672a\u6765\u72b6\u6001\u548c\u64cd\u4f5c\u52a8\u6001\u3002<\/p>\n

\u968f\u540e\uff0c\u6a21\u578b\u518d\u57fa\u4e8e\u8fd9\u4e9b\u9690\u7a7a\u95f4\u63a8\u7406 (latent reasoning) \u5e76\u884c\u751f\u6210action tokens\u3002<\/p>\n

\u8fd9\u4e00\u6b65\u89e3\u51b3\u7684\u662f\uff1a\u5982\u4f55\u8ba9\u52a8\u4f5c\u751f\u6210\u5efa\u7acb\u5728\u7269\u7406\u63a8\u7406\u4e4b\u4e0a\u3002<\/p>\n

\u7b2c\u4e8c\u9636\u6bb5\uff1aLAPO\u540c\u65f6\u4f18\u5316\u9690\u7a7a\u95f4\uff08latent\uff09\u548c\u52a8\u4f5c\uff08action\uff09<\/p>\n

LaST-R1\u7684\u6838\u5fc3\u7b97\u6cd5\u662fLAPO\uff08\u9690\u7a7a\u95f4\u5230\u52a8\u4f5c\u7b56\u7565\u4f18\u5316\uff0cLatent-to-Action Policy Optimization\uff09\u3002<\/p>\n

\u4f20\u7edf\u5177\u8eab\u5927\u6a21\u578bRL\u4e3b\u8981\u4f18\u5316action\uff0c\u800cLAPO\u5c06latent reasoning\u4e5f\u7eb3\u5165\u5f3a\u5316\u5b66\u4e60\u76ee\u6807\uff0c\u8ba9\u73af\u5883\u5956\u52b1\u540c\u65f6\u5851\u9020\u201c\u600e\u4e48\u60f3\u201d\u548c\u201c\u600e\u4e48\u52a8\u201d\u3002<\/p>\n

\u8bba\u6587\u4e2d\u6700\u5173\u952e\u7684\u662flatent-level ratio surrogate\uff1a<\/p>\n

\u5176\u4e2d\uff0c<\/p>\n

\u8868\u793arollout\u65f6\u65e7\u7b56\u7565\u751f\u6210\u7684latent sequence\uff0c<\/p>\n

\u8868\u793a\u5f53\u524d\u7b56\u7565\u91cd\u65b0\u751f\u6210\u7684latent sequence\uff0c<\/p>\n

\u63a7\u5236latent\u5206\u5e03\u5bbd\u5ea6\u3002<\/p>\n

\u76f4\u89c2\u6765\u8bf4\uff0c\u5982\u679c\u67d0\u6761\u8f68\u8ff9\u6210\u529f\uff0cLaST-R1\u4e0d\u4ec5\u4f1a\u5f3a\u5316\u5bf9\u5e94\u52a8\u4f5c\uff0c\u4e5f\u4f1a\u5f3a\u5316\u52a8\u4f5c\u4e4b\u524d\u4ea7\u751f\u7684\u201c\u597d\u63a8\u7406\u201d\u3002<\/p>\n

\u968f\u540e\uff0cLAPO\u5c06latent\u548caction\u653e\u8fdb\u7edf\u4e00\u7684clipped objective\u4e2d\uff1a<\/p>\n

\u5176\u4e2d\uff0c<\/p>\n

\u8868\u793a\u540c\u65f6\u4f18\u5316latent reasoning\u548caction generation\uff0c<\/p>\n

\u662fadvantage estimate\uff0c<\/p>\n

\u7528\u4e8e\u9650\u5236\u7b56\u7565\u66f4\u65b0\u5e45\u5ea6\u3002<\/p>\n

\u6700\u7ec8\uff0c\u603b\u8bad\u7ec3\u76ee\u6807\u4e3a\uff1a<\/p>\n

\u8fd9\u610f\u5473\u7740\uff1aLaST-R1\u7684RL\u540e\u8bad\u7ec3\u4e0d\u53ea\u662f\u4f18\u5316\u673a\u5668\u4eba\u7684\u52a8\u4f5c\u7ed3\u679c\uff0c\u4e5f\u5728\u4f18\u5316\u884c\u52a8\u524d\u7684\u7269\u7406\u63a8\u7406\u8fc7\u7a0b\u3002<\/p>\n

\u7b2c\u4e09\u9636\u6bb5\uff1aAdaptive Latent CoT<\/p>\n

\u4e0d\u540c\u4efb\u52a1\u9700\u8981\u4e0d\u540c\u7684\u601d\u8003\u957f\u5ea6\u3002<\/p>\n

\u56e0\u6b64\uff0cLaST-R1\u5f15\u5165Adaptive Latent CoT\uff0c\u901a\u8fc7token\u8ba9\u6a21\u578b\u52a8\u6001\u51b3\u5b9a\u4f55\u65f6\u7ed3\u675flatent reasoning\u5e76\u8fdb\u5165action generation\u3002<\/p>\n

\u8fd9\u662f\u4e3a\u4e86\u8ba9\u673a\u5668\u4eba\u6839\u636e\u4efb\u52a1\u96be\u5ea6\u81ea\u9002\u5e94\u5206\u914d\u63a8\u7406\u9884\u7b97\u3002<\/p>\n

\u4e5f\u5c31\u662f\u8bf4\uff0cLaST-R1\u4e0d\u662f\u8ba9\u673a\u5668\u4eba\u6bcf\u4e00\u6b65\u90fd\u56fa\u5b9a\u60f3\u540c\u6837\u4e45\uff0c\u800c\u662f\u8ba9\u5b83\u5b66\u4f1a\uff1a\u7b80\u5355\u72b6\u6001\u5feb\u901f\u6267\u884c\uff0c\u590d\u6742\u72b6\u6001\u591a\u60f3\u4e00\u6b65\u3002<\/p>\n

\u4e3a\u4e86\u4f18\u5316\u8fd9\u4e2a\u7ed3\u675f\u6807\u8bc6\u7b26token\u7684\u81ea\u9002\u5e94\u751f\u6210\uff0c\u8bad\u7ec3\u76ee\u6807\u9700\u8981\u8fdb\u4e00\u6b65\u52a0\u4e0aL_end\u3002<\/p>\n

1\u3001\u4eff\u771f\u5b9e\u9a8c\uff1aLIBERO 99.9%<\/p>\n

LaST-R1\u5728LIBERO benchmark\u4e0a\u8fdb\u884c\u7cfb\u7edf\u8bc4\u4f30\uff0c\u8986\u76d6Spatial\u3001Object\u3001Goal \u548cLong\u56db\u4e2a\u4efb\u52a1\u5957\u4ef6\u3002\u5b9e\u9a8c\u5728one-shot SFT warm-up\u8bbe\u7f6e\u4e0b\u8fdb\u884c\uff0c\u968f\u540e\u8fdb\u5165\u5728\u7ebfRL\u540e\u8bad\u7ec3\u3002<\/p>\n

\u7ed3\u679c\u663e\u793a\uff0cLaST-R1\u5728\u56db\u4e2asuite\u4e0a\u5206\u522b\u8fbe\u523099.8%\/100.0%\/100.0%\/99.8%\uff0c\u5e73\u5747\u6210\u529f\u7387\u8fbe\u523099.9%\uff0c\u8d85\u8fc7OpenVLA-OFT\u3001\u03c00.5\u3001SimpleVLA-RL\u548c\u03c0RL\u7b49\u5f3a\u57fa\u7ebf\u3002<\/p>\n

\u76f8\u6bd4\u53ea\u4f18\u5316\u52a8\u4f5c\u7a7a\u95f4\u7684Action-Only + PPO\uff0cLaST-R1 + LAPO\u6536\u655b\u66f4\u5feb\u3001\u6700\u7ec8\u6210\u529f\u7387\u66f4\u9ad8\uff0c\u8bf4\u660elatent reasoning\u4e0eaction generation\u7684\u8054\u5408\u4f18\u5316\u80fd\u591f\u4e3aRL\u63d0\u4f9b\u66f4\u7a33\u5b9a\u7684\u201c\u8ba4\u77e5\u7f13\u51b2\u533a\u201d\uff0c\u4ece\u800c\u63d0\u5347\u590d\u6742\u957f\u7a0b\u64cd\u4f5c\u80fd\u529b\u3002<\/p>\n

2\u3001\u771f\u673a\u5b9e\u9a8c\uff1a\u4ece52.5%\u523093.75%<\/p>\n

LaST-R1\u5728\u56db\u4e2a\u771f\u5b9e\u64cd\u4f5c\u4efb\u52a1\u4e0a\u8fdb\u884c\u6d4b\u8bd5\uff0c\u8986\u76d6\u5355\u81c2\u9ad8\u7cbe\u5ea6\u63d2\u5165\u3001\u53cc\u81c2\u534f\u540c\u3001\u63a5\u89e6\u5f0f\u64e6\u62ed\u548c\u8fde\u7eed\u65cb\u8f6c\u7b49\u590d\u6742\u7269\u7406\u4ea4\u4e92\u3002<\/p>\n

\u4e3a\u4e86\u7a81\u51faRL\u540e\u8bad\u7ec3\u6548\u679c\uff0c\u8bba\u6587\u5c06\u5176\u4e0eSOTA\u6a21\u578b\u03c00.5\u5bf9\u6bd4\uff1a\u03c00.5\u4f7f\u7528100\u6761\u4e13\u5bb6\u8f68\u8ff9\u8fdb\u884cSFT\uff0c\u800cLaST-R1\u4ec5\u4f7f\u752830\u6761\u8f68\u8ff9warm-up\uff0c\u5e76\u901a\u8fc7RL\u540e\u8bad\u7ec3\u7ee7\u7eed\u4f18\u5316\u3002<\/p>\n

\u7ed3\u679c\u663e\u793a\uff0cLaST-R1\u5c06\u771f\u673a\u5e73\u5747\u6210\u529f\u7387\u4ecewarmup\u540e\u768452.5%\u63d0\u5347\u523093.75%\uff0c\u663e\u8457\u8d85\u8fc7\u03c00.5\u768471.25%\uff0c\u8bf4\u660e\u5176\u4f18\u52bf\u4e0d\u4ec5\u5b58\u5728\u4e8e\u4eff\u771f\u73af\u5883\uff0c\u4e5f\u80fd\u8fc1\u79fb\u5230\u771f\u5b9e\u7269\u7406\u4ea4\u4e92\u4e2d\uff0c\u5e76\u5f62\u6210\u66f4\u7a33\u5b9a\u7684\u6267\u884c\u7b56\u7565\u3002<\/p>\n

3\u3001\u6cdb\u5316\u5b9e\u9a8c\uff1a\u6362\u7269\u4f53\u3001\u6362\u80cc\u666f\u3001\u6362\u5149\u7167\uff0c\u4f9d\u7136\u7a33<\/p>\n

\u5728LIBERO OOD\u8bbe\u7f6e\u4e2d\uff0c\u7814\u7a76\u56e2\u961f\u91c7\u75289\u4e2aseen tasks\u8fdb\u884c\u5728\u7ebfRL\uff0c\u5e76\u4fdd\u75591\u4e2aheld-out task\u505a\u6cdb\u5316\u6d4b\u8bd5\u3002<\/p>\n

\u7ed3\u679c\u663e\u793a\uff0cAction-Only + PPO\u5bb9\u6613\u51fa\u73b0\u6027\u80fd\u505c\u6ede\u751a\u81f3\u9000\u5316\uff0c\u800cLaST-R1 + LAPO\u80fd\u5728OOD tasks\u4e0a\u6301\u7eed\u63d0\u5347\uff0c\u8bf4\u660elatent reasoning\u80fd\u5e2e\u52a9\u6a21\u578b\u5b66\u5230\u66f4\u53ef\u8fc1\u79fb\u7684\u7a7a\u95f4\u8bed\u4e49\u548c\u7269\u7406\u52a8\u6001\u3002<\/p>\n

\u5728\u771f\u5b9e\u4e16\u754c\u4e2d\uff0c\u8bba\u6587\u8fdb\u4e00\u6b65\u6d4b\u8bd5\u4e86unseen object\u3001background variation\u548clighting condition\u4e09\u7c7b\u6270\u52a8\u3002<\/p>\n

\u76f8\u6bd4SFT \u03c00.5\uff0cLaST-R1\u5728\u8fd9\u4e9b\u53d8\u5316\u4e0b\u4fdd\u6301\u66f4\u5c0f\u7684\u6027\u80fd\u4e0b\u964d\uff0c\u8bf4\u660e\u5b83\u5e76\u4e0d\u662f\u7b80\u5355\u8bb0\u4f4f\u8bad\u7ec3\u573a\u666f\u4e2d\u7684\u52a8\u4f5c\u8f68\u8ff9\uff0c\u800c\u662f\u5f62\u6210\u4e86\u66f4\u9c81\u68d2\u7684\u7269\u7406\u63a8\u7406\u4e0e\u52a8\u4f5c\u751f\u6210\u80fd\u529b\u3002<\/p>\n

LaST-R1\u7684\u610f\u4e49\uff0c\u4e0d\u53ea\u662f\u628aLIBERO\u5e73\u5747\u6210\u529f\u7387\u63a8\u523099.9%\uff0c\u4e5f\u4e0d\u53ea\u662f\u8ba9\u771f\u673a\u4efb\u52a1\u6210\u529f\u7387\u63d0\u5347\u523093.75%\u3002<\/p>\n

\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u5177\u8eab\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u8303\u5f0f\uff1a\u5f3a\u5316\u5b66\u4e60\u4e0d\u5e94\u8be5\u53ea\u4f18\u5316\u673a\u5668\u4eba\u7684\u52a8\u4f5c\uff0c\u4e5f\u5e94\u8be5\u4f18\u5316\u52a8\u4f5c\u80cc\u540e\u7684\u7269\u7406\u63a8\u7406\u8fc7\u7a0b\u3002<\/p>\n

\u8fc7\u53bb\uff0c\u6211\u4eec\u66f4\u5173\u5fc3\u673a\u5668\u4eba\u80fd\u4e0d\u80fd\u751f\u6210\u6b63\u786e\u52a8\u4f5c\u3002<\/p>\n

\u73b0\u5728\uff0cLaST-R1\u5728\u6b64\u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u8ffd\u95ee\uff1a\u673a\u5668\u4eba\u80fd\u4e0d\u80fd\u5728\u884c\u52a8\u524d\u8fdb\u884c\u6b63\u786e\u7684\u7269\u7406\u63a8\u7406\uff1f<\/p>\n

\u901a\u8fc7LAPO\uff0c\u73af\u5883reward\u53ef\u4ee5\u76f4\u63a5\u5851\u9020latent reasoning space\uff1b<\/p>\n

\u901a\u8fc7adaptive latent CoT\uff0c\u673a\u5668\u4eba\u53ef\u4ee5\u6839\u636e\u4efb\u52a1\u96be\u5ea6\u52a8\u6001\u8c03\u6574\u601d\u8003\u957f\u5ea6\u3002<\/p>\n

\u8fd9\u610f\u5473\u7740\uff0c\u673a\u5668\u4eba\u4e0d\u518d\u53ea\u662f\u590d\u73b0\u6f14\u793a\u6570\u636e\u4e2d\u7684\u52a8\u4f5c\u8f68\u8ff9\uff0c\u800c\u662f\u5728\u4ea4\u4e92\u4e2d\u9010\u6b65\u5f3a\u5316\u6a21\u578b\u7684\u7269\u7406\u63a8\u7406\u3002<\/p>\n

\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u770b\uff0cLaST-R1\u8ba9\u5177\u8eab\u5927\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\u4ece\u201c\u770b\u89c1\u5c31\u52a8\u201d\u8d70\u5411\u201c\u5148\u60f3\u660e\u767d\uff0c\u518d\u7a33\u5b9a\u884c\u52a8\u201d\u3002<\/p>\n

\u5f53\u5177\u8eab\u5927\u6a21\u578b\u5f00\u59cb\u5b66\u4f1a\u5728latent space\u4e2d\u601d\u8003\uff0c\u673a\u5668\u4eba\u8ddd\u79bb\u771f\u6b63\u7684\u81ea\u4e3b\u64cd\u4f5c\uff0c\u4e5f\u8bb8\u53c8\u8fd1\u4e86\u4e00\u6b65\u3002<\/p>\n

\u8bba\u6587\u94fe\u63a5: https:\/\/arxiv.org\/abs\/2604.28192\u9879\u76ee\u4e3b\u9875: https:\/\/siriyep.github.io\/last-r1\/\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/CHEN-H01\/LaST-R1<\/p>\n

\n

\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n

\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n

\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n

\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

\u5141\u4e2d \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u673a\u5668\u4eba\u62c9\u4e2a\u62c9\u94fe\uff0c\u5230\u5e95\u9700\u4e0d\u9700\u8981\u201c\u8111\u5b50\u201d\uff1f<\/p>\n

\u8fc7\u53bb\u51e0\u5e74\uff0c\u4eceOpenVLA\u5230\u03c00\u3001\u03c00.5\uff0c\u5177\u8eab\u5927\u6a21\u578b\u5df2\u7ecf\u80fd\u8ba9\u673a\u5668\u4eba\u628a\u6307\u4ee4\u548c\u52a8\u4f5c\u8fde\u5f97\u6709\u6a21\u6709\u6837\u3002<\/p>\n

\u4f46\u4e00\u65e6\u5305\u7684\u4f4d\u7f6e\u632a\u4e86\u51e0\u5398\u7c73\uff0c\u6216\u8005\u5149\u7167\u6697\u4e86\u4e00\u70b9\uff0c\u5b83\u4eec\u5f80\u5f80\u5c31\u4f1a\u201c\u5927\u8111\u5b95\u673a\u201d\u3002<\/p>\n

\u7a76\u5176\u539f\u56e0\uff0c\u662f\u56e0\u4e3a\u8fd9\u4e9b\u673a\u5668\u4eba\u5927\u591a\u5728\u73a9\u201c\u8fde\u8fde\u770b\u201d\uff1a\u770b\u5230\u89c2\u5bdf\u7ed3\u679c\uff0c\u76f4\u63a5\u8f93\u51fa\u52a8\u4f5c\u3002<\/p>\n

\u5b83\u4eec\u53ea\u662f\u8bb0\u4f4f\u4e86\u8f68\u8ff9\uff0c\u5374\u5e76\u4e0d\u7406\u89e3\u80cc\u540e\u7684\u7269\u7406\u903b\u8f91\u3002<\/p>\n

\u73b0\u5728\uff0c\u4e00\u79cd\u8ba9\u673a\u5668\u4eba\u201c\u5148\u60f3\u660e\u767d\uff0c\u518d\u7a33\u5b9a<\/p>\n","protected":false},"author":1,"featured_media":4663,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[800,8,801,802],"class_list":["post-4667","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-last-r1","tag-8","tag-801","tag-802"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/4667","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=4667"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/4667\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/4663"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=4667"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=4667"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=4667"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}